Hammerspace AI storage pitch targets GPU data bottlenecks

Hammerspace is positioning its Data Platform as hammerspace ai storage infrastructure for companies whose GPUs are waiting on data rather than compute. The company argues that the AI infrastructure constraint has shifted from raw storage capacity to where data sits, how it is accessed and how often teams copy it between systems.

The platform sits between compute and existing storage, including NAS, object stores and NVMe drives inside GPU servers, according to Hammerspace. It presents that storage through a single global namespace and supports standard protocols such as NFS, SMB and S3, which the company says avoids proprietary clients and application rewrites. Pricing, revenue and customer counts were not disclosed.

What is the Hammerspace Data Platform?

The Hammerspace Data Platform is a data orchestration layer that makes files across different storage systems, sites and clouds appear in one namespace. Its purpose is to let applications find and use data where it already resides, while policies decide when copies should move to faster tiers.

Hammerspace’s argument starts with fragmentation. Jonathan Flynn, director of applied systems at Hammerspace, said enterprise training data is often spread across departments, business units and storage silos rather than prepared as a clean training set. Gartner has said 57% of organizations believe their data is not AI-ready, and that two-thirds of executives believe no one in their organization understands all the data they have collected and how to access it.

Mike Bloom, who covers AR architecture at Hammerspace, said the common vendor answer of replacing legacy arrays misses where enterprise data actually lives. Hammerspace instead describes its approach as assimilation: scan an existing NAS, ingest the directory tree into the global namespace and redirect mounts while the underlying system, such as NetApp, Qumulo or VAST, keeps serving the bytes.

How Hammerspace uses server NVMe

A second part of the pitch is the NVMe capacity already installed in GPU servers. Hammerspace says HGX and DGX systems typically ship with eight to sixteen NVMe drives, but orchestration layers often treat that capacity as local scratch space for one machine. The company calls that stranded storage and says it can amount to hundreds of terabytes per server, with two-petabyte GPU servers on the roadmap.

Hammerspace calls the resulting shared layer Tier 0. Flynn said this can be faster and cheaper than buying a separate all-flash array because the compute nodes, network and drives are already in the rack. He contrasted GPU-server NVMe lanes with traditional 2U storage appliances that may expose far fewer network lanes than internal PCIe lanes.

The company says policy objectives can move hot data to Tier 0 without users running copy or rsync jobs. Bloom said the same mechanism can remove those copies after a training run, which helps keep the hot tier from congesting and affecting other workloads.

Benchmarks and open standards claims

Hammerspace is leaning on standards as part of its AI infrastructure claim. A Samsung-Hammerspace submission placed in the top 10 of the IO500 10-Node Production benchmark in November 2025, using standard Linux, the upstream NFSv4.2 client, standard NVMe SSDs and IP-over-InfiniBand, according to the company.

The company also submitted MLPerf Storage v2.0 results showing scaling to 420.8 GB/s across 140 GPUs on five nodes, with GPU utilization above 96%. Those are vendor-submitted results, and buyers would still want independent comparisons against VAST, WekaIO and NetApp in mixed customer environments.

Under the hood, Hammerspace uses parallel NFS, or pNFS, rather than traditional single-server NFS. pNFS became RFC 5661 in 2010, and Hammerspace says it helped extend NFSv4.2 in 2018 through the Flex Files extension, which supports heterogeneous storage across cloud tiers, legacy file systems and multiple sites.

The platform’s policy layer also covers data location rules, according to Hammerspace. A dataset tagged EU-only can be excluded from North American volumes, while HIPAA-bound data can be governed with write-once-read-many rules. Hammerspace says Meta uses the platform across two 24,576-GPU clusters used to train Llama 3, specifically for live job debugging and real-time code propagation across training pipelines.

This story draws on original reporting from The Register.